Why do AI agents fail to execute on Solana?

Short answer: Most AI agent payments today are sub-cent API calls, and the rails for those work. The harder problem appears when an agent trades: a sound, approved decision can still be delayed or outbid on the way to the block, and an agent that can't tell whether it landed will retry. On Solana, whether a transaction lands is decided by where the software runs relative to the validator building the block.

Key takeaways

BeInCrypto's State of AI Agent Payments 2026 finds that 90.8% of recorded agent payments were worth less than one cent, and that institutional financial workflows scored 12.0 out of 100 on readiness.
Solana handled 94.5% of the transactions in the report.
For trading agents, the bottleneck isn't the wallet. It's the gap between the decision and the confirmed trade.
Solana slots have been ~200ms since October 2026, and Alpenglow targets ~150ms finality from 21 October 2026. Faster blocks make timing more precise, not less important.
Solana co-location means running trading software on the same hardware as the validator that builds the block. That's what "trade where the block is built" means.

What did BeInCrypto find about AI agent payments?

On 14 October 2026, BeInCrypto published "The Uncomfortable Truth Behind 90% of AI Payments" by Brian McGleenon. It summarises the outlet's report The State of AI Agent Payments 2026, produced with AWS, Cloudflare, Chainalysis and Zoomex. The report studied payments on x402, a payment protocol designed for machines. Its main findings (BeInCrypto (https://beincrypto.com/uncomfortable-truth-ai-agent-payments-2026/)):

90.8% of recorded agent payments were microtransactions worth less than one cent.
A 35-day sample of 6.4 million transactions moved just $119,947 in total.
On the report's readiness framework, micropayments scored 60.7/100 and institutional financial workflows scored 12.0/100.
14 of 19 trading tools and financial systems reviewed operate only within human-set spending limits.
AI-labelled tools were under 0.55% of trading volume on Hyperliquid.
$281,200 was lost to exploits that redirected validly budgeted capital.
Solana handled 94.5% of the transactions in the report.
92% ($3.91B) of VC funding went to payment and cash-management infrastructure, not trading bots.

In short, the payment rails work, but demand is early and humans still sign off on anything that matters.

Why is trading harder for AI agents than paying for an API?

Paying for an API is a one-way transfer of a fixed amount. Missing the next block costs almost nothing. A trade is different: its value depends on the moment it executes. If it lands late, the price has moved. If it's outbid, the opportunity has gone.

Robin Nordnes, Raiku's founder and CEO, described the risk in the BeInCrypto piece as an execution gap:

"The risk I worry about more sits between the decision and the confirmed trade. Say an agent makes a sound, compliant trade within budget, and the transaction gets delayed or outbid on the way to the block. If the agent can't tell whether it went through, it tries again. Then again. Every retry was individually approved, and together they can do real damage."

Spending limits and wallet permissions control what an agent may do. They don't control whether its transaction lands. That's why the report's institutional readiness score is so low. A desk can't hand over execution until execution behaves predictably.

What decides whether a transaction lands on Solana?

On Solana, one validator, the leader, builds each block for its slot. Every transaction has to reach that leader in time. Software that runs far from the leader, across the public internet and through relays, spends part of every slot just travelling.

Slot time: Solana cut slot times from 400ms to 200ms in four stages, completed in October 2026 (solana.com (https://solana.com/upgrades/alpenglow)).
Finality: Alpenglow, scheduled to activate in epoch 1064 on 21 October 2026, targets roughly 150ms finality, down from about 12.8 seconds. Solana describes 150ms as "an expectation, not a constant" (solana.com (https://solana.com/upgrades/alpenglow); see also CoinDesk (https://www.coindesk.com/tech/2026/10/14/solana-sets-oct-21-for-alpenglow-upgrade-targeting-150-millisecond-finality)).

Faster finality helps agents a lot: they'll know sooner whether a trade is final, which cuts the blind retries Nordnes describes. But shorter slots also shrink the window to get a trade in. A fixed delay, the distance between the software and the block, becomes a larger share of every slot.

What is Solana co-location?

Solana co-location means running trading software on the same hardware as the validator that builds the block. It's the Solana version of a familiar idea from traditional markets, where firms place servers beside an exchange. On Solana, the venue is the validator producing each block, so co-location means being on that machine, not next to it.

Raiku Blackline applies this idea. A firm's strategy runs on the same hardware as Raiku's custom validator. As raiku.com describes it, the process has three steps (raiku.com (https://raiku.com)):

See: the validator sees each transaction as it executes, and the strategy sees it microseconds later.
Act: the strategy acts before the block reaches the rest of the network.
Land: its transactions are included in blocks Raiku produces. Raiku does not set the price or guarantee the outcome.

The strategy and its profits stay with the firm. Blackline is built for a small number of trading firms and asset managers, and access is by invitation. In short: trade where the block is built.

What does this mean for AI trading agents?

Raiku doesn't build AI agents or payment rails. But any automated strategy faces the same physics, whether a quant wrote it or a model generates its decisions. Three practical points for desks considering agent-driven execution on Solana:

Separate decision risk from execution risk. Guardrails govern the first. Infrastructure governs the second.
Design for finality after Alpenglow. Solana advises pointing new code at `finalized` once Alpenglow is live, and not before activation (solana.com (https://solana.com/upgrades/alpenglow)).
Measure where your software runs. In 200ms slots, location is part of the strategy, not an IT detail.

FAQ

Are AI agents making real payments on crypto rails?

Yes, but they're mostly tiny. BeInCrypto's report found that 90.8% of recorded agent payments were under one cent, and a 35-day sample of 6.4 million transactions totalled $119,947.

Why do AI agents fail to execute trades on Solana?

A correct, approved trade can be delayed or outbid before it reaches the block. If the agent can't confirm the outcome, repeated retries can compound losses.

What is Solana co-location?

Running trading software on the same hardware as the validator that builds the block, so the software sees executed transactions sooner and has less distance to cover before the block is built.

How fast is Solana in 2026?

Slot times are ~200ms (reduced from 400ms in stages completed in October 2026). Alpenglow targets ~150ms finality from 21 October 2026, down from about 12.8 seconds.

Does Alpenglow solve the execution problem?

It reduces uncertainty about whether a trade is final. It doesn't change how far a transaction has to travel to reach the block producer, and in shorter slots that distance matters more.

Who can use Raiku Blackline?

Blackline is built for a small number of trading firms and asset managers, by invitation and subject to a formal agreement. Contact Raiku via raiku.com.